ChatGPT Quality, Accuracy and Brand Risk for Marketers

#01Where do errors come from?

#02Which data settings matter?

• Personal plans: "Improve the model for everyone" may be on by default. Switch it off in Settings → Data controls if you upload work material.
• Business, Enterprise and Edu: OpenAI says it doesn't train on your data by default; admins control apps, plugins and Work access.
• Temporary chat: doesn't appear in history and isn't used to train models.
• Apps and plugins: permissions decide when ChatGPT asks before reading or acting; installation can't bypass them.

Check your company's own policy too. Many forbid uploading customer personal data to any external tool.

#03What should every review check?

PRE-PUBLISH REVIEW
Review this [asset] before publication. Return a table:
1) every factual claim and its source (or "none");
2) any comparison with Competitors A-D and its source;
3) anything that could read as a guarantee or promise;
4) anything off-voice per the project rules;
5) personal data or confidential details.
Don't rewrite. Just flag.

Our view: the review prompt is useful, but it isn't the review. A person who knows the product signs off anything public.

#04What should a team AI policy say?

Keep it to one page:

• Which plans and workspaces are approved, and the data settings to use.
• What may be uploaded (de-identified exports, public docs) and what may not (customer personal data, unreleased financials).
• Which outputs need human review (all public content, all ads, all customer emails).
• What's never automated (sending, posting, publishing, spending).
• How to disclose AI use where a platform or regulator requires it.
• Who owns the prompt library and skills.

#05Worked example: Ledgerline's policy

Ledgerline is a fictional accounts-payable tool. Details are illustrative.

Ledgerline moved its marketing team from personal Plus accounts to a Business workspace so data wasn't used for training by default and admins could control plugins. The one-page policy banned uploading customer invoices, required review of every public claim and kept sending in human hands. In the first month, the pre-publish review flagged an unsourced "most popular" claim in a landing page, which was removed.

#06What should you check before the next lesson?

• Check Data controls on every account your team uses.
• Write a one-page AI policy using the list above.
• Add the pre-publish review to your workflow.
• Compare with Claude marketing governance and privacy and Claude editing and fact-checking.

Next: ChatGPT search: how answers cite brands. Or return to the ChatGPT for Marketers track.

#07Common questions

Does ChatGPT train on my marketing data? On personal Free, Go, Plus and Pro accounts, "Improve the model for everyone" can be on, and you can switch it off in Data controls. OpenAI says it does not train on Business, Enterprise or Edu data by default.

How do I stop ChatGPT inventing facts? Give it a fact sheet, tell it to use only those facts, ask it to mark unsupported claims and check citations yourself. No setting removes errors entirely.

Can we use ChatGPT for political or health marketing? OpenAI's usage policies restrict political campaigning and some sensitive uses. Check the current policies and your own regulator's rules first.